Hacker News Reader

lanceflton At 14 Justine Tunney published software. Now she makes AI models run from a file

Remember that she renamed the GGML format to "GGJT" after her own name, and lied about mmap memory reduction after stealing the code from another user.

i.e. she said mmap reduced it 90% but it's how Linux displays it, actually similar memory is needed.

Then she was banned from llama.cpp and had to make her own fork/imitation.

lanceflton Mondrian Entered the Public Domain. The Estate Disagrees

Council taxes are property taxes and are monthly.

lanceflton Launch HN: Risely (YC S25) – AI Agents for Universities

Thanks! I've emailed.

lanceflton Launch HN: Risely (YC S25) – AI Agents for Universities

The key issue for the sector is the tens of legacy systems that don't integrate with each other, often with manual spreadsheet processes that could be easily automated. Yet the big players like Oracle sell a generic CRM experience that doesn't fit well with higher education.

Are you hiring? I have 8 years of university SIS implementation & migration experience and 2 years of Edtech AI engineering experience and this is the exact problem space I want to work in.

lanceflton Apple M3 Ultra

Deepseek R1 (full, Q1) is 14t/s on an M2 Ultra, so this should be around 20t/s

lanceflton Tencent Hunyuan-Large

RAM for 4-bit is 1GB per 2 billion parameters. So you will want 256GB RAM and at least one GPU. If you only have one server and one user, it's the full parameter count. (If you have multiple GPUs/servers and many users in parallel, you can shard and route it so you only need the active parameter count per GPU/server. So it's cheaper at scale.)

lanceflton Leak claims RTX 5090 has 600W TGP, RTX 5080 hits 400W

I'm running a 4090 at 280W, and I'm seeing ~96% of the performance of 450W. There's no need to run it at full power.

lanceflton Have Swiss scientists made a chocolate breakthrough?

This is just an ad for the Swiss chocolate industry. The only people quoted are being funded directly by chocolate manufacturers.

lanceflton Extracting concepts from GPT-4

https://github.com/openai/sparse_autoencoder

They actually open sourced it, for GPT-2 which is an open model.

lanceflton Llama 3-V: Matching GPT4-V with a 100x smaller model and 500 dollars

They didn't train the vision encoder either, it's unchanged SigLIP by Google.

lanceflton Llama 3-V: Matching GPT4-V with a 100x smaller model and 500 dollars

- Llava is not the SOTA open VLM, InternVL-1.5 is https://huggingface.co/spaces/opencompass/open_vlm_leaderboa...

You need to compare the evals to strong open VLMs including this and CogVLM

- This is not "first-ever multimodal model built on top of Llama3", there's already a Llava on Llama3-8b https://huggingface.co/lmms-lab